Remote Sensing Image De-striping using Deep Convolutional Networks

Optical remote sensing images are affected by stripe noise due to the inconsistent photo response among detector elements and make the structure of ground features hard to recognize and affects the subsequent higher level data applications and geo-physical parameter retrievals. Image destriping is a vital pre-processing technique in remote sensing aimed to enhance the quality and interpretability of acquired imagery. Existing de-striping algorithms struggle to balance noise suppression, detail preservation and turnaround time. SNRWDNN (Stripe Noise Removal Wavelet Deep Neural Network) is deep convolutional network takes the advantage of wavelets in suppressing the stripe noise is explored on real stripe images. This paper discusses the characterization of stripes in IRS-1C LISS-III sensor data and training the network model based on the stripe characteristics. Performance of this model is evaluated with well-known optimization-based models such as LRSID (Low-Rank Single-Image Decomposition) and directionally non-convex $\mathrm{l}_{0}$ sparse model. The model outperforms other methods in effectively removing the stripes without losing sharpness of the images with 20-fold improvement in turnaround time.